Research talk: Project Dexter: Machine learning and automatic decision-making for robotic manipulation
Robot technology has long held the promise of disrupting many important industries that involve dexterous object manipulation in weakly structured environments, including healthcare, agriculture, and infrastructure maintenance. The increasing versatility of robotic manipulation hardware seemingly puts this disruption within reach. However, hardware versatility comes at a cost. Its complexity defies traditional control-based approaches, and recent success stories, such as a dexterous robotic hand assembling a Rubik’s cube, highlight the reality: it can take world-class roboticists dozens of person-years to train an advanced robotic manipulator to solve a single problem with modern machine learning–based sequential decision-making techniques. The goal of Microsoft Research’s Project Dexter is to enable training robotic manipulation policies for real-world tasks at a practical cost in terms of expertise, time, and compute. This talk will give an overview of Project Dexter’s research agenda, which focuses on learning generalizable representations of visual, tactile, and proprioceptive observation data on Transformer-based architectures, and on both reinforcement learning and imitation learning approaches capable of using this generalized knowledge with other prior information to achieve practically acceptable sample efficiency. We will outline several directions we are currently exploring in this large problem space and briefly delve into one of them—an offline reinforcement learning approach the team is researching in the robotic manipulation context.
Learn more about the 2021 Microsoft Research Summit: https://Aka.ms/researchsummit (opens in new tab)
- 轨迹:
- Reinforcement Learning
- 日期：
- 演讲者：
- Andrey Kolobov, Ching-An Cheng
- 所属机构：
- Microsoft Research Redmond
-
-
Andrey Kolobov
Principal Research Manager
-
Ching-An Cheng
Senior Researcher
-
-
Reinforcement Learning
-
-
-
Research talk: Reinforcement learning with preference feedback
Speakers:- Aadirupa Saha
-
-
-
-
-
-
Panel: Generalization in reinforcement learning
Speakers:- Mingfei Sun,
- Roberta Raileanu,
- Harm van Seijen
-
-
Research talk: Successor feature sets: Generalizing successor representations across policies
Speakers:- Kiante Brantley
-
Research talk: Towards efficient generalization in continual RL using episodic memory
Speakers:- Mandana Samiei
-
Research talk: Breaking the deadly triad with a target network
Speakers:- Shangtong Zhang
-
Panel: The future of reinforcement learning
Speakers:- Geoff Gordon,
- Emma Brunskill,
- Craig Boutilier
-
